CN109635932A - A kind of Graphic Design and implementation method of part connection multilayer perceptron - Google Patents
A kind of Graphic Design and implementation method of part connection multilayer perceptron Download PDFInfo
- Publication number
- CN109635932A CN109635932A CN201811538198.3A CN201811538198A CN109635932A CN 109635932 A CN109635932 A CN 109635932A CN 201811538198 A CN201811538198 A CN 201811538198A CN 109635932 A CN109635932 A CN 109635932A
- Authority
- CN
- China
- Prior art keywords
- network
- multilayer perceptron
- layer
- capsule
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 28
- 238000013461 design Methods 0.000 title claims abstract description 6
- 238000012549 training Methods 0.000 claims abstract description 44
- 210000002569 neuron Anatomy 0.000 claims abstract description 18
- 238000004364 calculation method Methods 0.000 claims abstract description 15
- 239000002775 capsule Substances 0.000 claims description 39
- 239000011159 matrix material Substances 0.000 claims description 31
- 230000004913 activation Effects 0.000 claims description 23
- 238000012360 testing method Methods 0.000 claims description 21
- 238000013528 artificial neural network Methods 0.000 claims description 9
- 230000006870 function Effects 0.000 claims description 9
- 206010012601 diabetes mellitus Diseases 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 7
- 241001164374 Calyx Species 0.000 claims description 4
- 230000007423 decrease Effects 0.000 claims description 4
- 239000008280 blood Substances 0.000 claims description 3
- 210000004369 blood Anatomy 0.000 claims description 3
- 238000013480 data collection Methods 0.000 claims description 2
- 239000003292 glue Substances 0.000 claims 1
- 210000002364 input neuron Anatomy 0.000 claims 1
- 210000005036 nerve Anatomy 0.000 claims 1
- 230000000694 effects Effects 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- NOESYZHRGYRDHS-UHFFFAOYSA-N insulin Chemical compound N1C(=O)C(NC(=O)C(CCC(N)=O)NC(=O)C(CCC(O)=O)NC(=O)C(C(C)C)NC(=O)C(NC(=O)CN)C(C)CC)CSSCC(C(NC(CO)C(=O)NC(CC(C)C)C(=O)NC(CC=2C=CC(O)=CC=2)C(=O)NC(CCC(N)=O)C(=O)NC(CC(C)C)C(=O)NC(CCC(O)=O)C(=O)NC(CC(N)=O)C(=O)NC(CC=2C=CC(O)=CC=2)C(=O)NC(CSSCC(NC(=O)C(C(C)C)NC(=O)C(CC(C)C)NC(=O)C(CC=2C=CC(O)=CC=2)NC(=O)C(CC(C)C)NC(=O)C(C)NC(=O)C(CCC(O)=O)NC(=O)C(C(C)C)NC(=O)C(CC(C)C)NC(=O)C(CC=2NC=NC=2)NC(=O)C(CO)NC(=O)CNC2=O)C(=O)NCC(=O)NC(CCC(O)=O)C(=O)NC(CCCNC(N)=N)C(=O)NCC(=O)NC(CC=3C=CC=CC=3)C(=O)NC(CC=3C=CC=CC=3)C(=O)NC(CC=3C=CC(O)=CC=3)C(=O)NC(C(C)O)C(=O)N3C(CCC3)C(=O)NC(CCCCN)C(=O)NC(C)C(O)=O)C(=O)NC(CC(N)=O)C(O)=O)=O)NC(=O)C(C(C)CC)NC(=O)C(CO)NC(=O)C(C(C)O)NC(=O)C1CSSCC2NC(=O)C(CC(C)C)NC(=O)C(NC(=O)C(CCC(N)=O)NC(=O)C(CC(N)=O)NC(=O)C(NC(=O)C(N)CC=1C=CC=CC=1)C(C)C)CC1=CN=CN1 NOESYZHRGYRDHS-UHFFFAOYSA-N 0.000 description 2
- 238000003909 pattern recognition Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- WQZGKKKJIJFFOK-GASJEMHNSA-N Glucose Natural products OC[C@H]1OC(O)[C@H](O)[C@@H](O)[C@@H]1O WQZGKKKJIJFFOK-GASJEMHNSA-N 0.000 description 1
- 102000004877 Insulin Human genes 0.000 description 1
- 108090001061 Insulin Proteins 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000036772 blood pressure Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 239000008103 glucose Substances 0.000 description 1
- 229940125396 insulin Drugs 0.000 description 1
- 150000002632 lipids Chemical class 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 230000035935 pregnancy Effects 0.000 description 1
- 210000000434 stratum corneum Anatomy 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses the Graphic Design and implementation method of a kind of part connection multilayer perceptron, selection needs the data set of training, meets data set data prediction the needs of network training;Corresponding data structure is generated according to the network graphic that user draws, and checks whether the connection type of network is legal;It sorts to network structural topology, the forward calculation of network is executed according to linear order, calculates the output valve x on neuron;According to the retrospectively calculate for the linear order reverse execution network that topological sorting obtains, the error amount δ on neuron is calculated;Directed edge in traverses network calculates the weighted value w on gradient updating directed edge;The present invention can be used to the multilayer perceptron of training part connection, and generate network by graphic method, without generating network by programming.
Description
Technical field
The present invention relates to the Graphic Designs and implementation method of a kind of part connection multilayer perceptron, belong to pattern-recognition
Field.The present invention can be used to train the multilayer perceptron of various connection ratios, and generate artificial mind by graphic method
Through network rather than traditional programming is realized.
Background technique
As a kind of technology for realizing artificial intelligence, artificial neural network is in image recognition, speech recognition, natural language
The multiple fields such as speech processing achieve breakthrough achievement, and receive very high concern.Multilayer perceptron was once suitable
Popular machine learning algorithm, possesses and is widely applied scene.In recent years, due to the success of deep learning, multilayer perceptron is again heavy
Newly paid close attention to.
Multilayer perceptron is a kind of feedforward neural network by there is the digraph of multilayer node to constitute, and wherein each is non-defeated
Ingress is all the neuron with nonlinear activation function, and each layer is connect entirely with next layer.The one of multilayer perceptron
A advantage is that it can be trained to.The process of training multilayer perceptron is exactly optimization mathematically in fact, and target is to find one
The optimal weight of group reduces the error of predicted value and actual value to greatest extent.
A given training sample is indicated with (x, y), calculates the activation value on each neuron first, to the last one
Layer.After the last layer calculates, the error of actual value and predicted value is calculated.Further according to obtained error retrospectively calculate
The error delta value on each neuron on each layer, the value can be understood as the neuron for the value of overall error,
The calculating process of error is promoted from the last layer.It is finally undated parameter, target is that error is made to become smaller.
Using element product (Hadamardproduct) is indicated, the matrix representation forms of multilayer perceptron algorithm are as follows
It is shown:
1. executing forward calculation process, the second layer is calculated up to the activation value of the last layer, subscript n is indicated more here
The n-th layer of layer perceptron, nlIndicate the last layer, that is, output layer of multilayer perceptron.
2. first finding out the error amount of the last layer, y indicates that the label value of sample, z indicate a upper layer data as input warp
The value after weighted sum is crossed, a indicates value of the z after activation primitive activation,Indicate that sample passes through multilayer perceptron
Predicted value after calculating, f'(z) indicate f (z) derivative:
3. inverted order seeks the error amount from layer second from the bottom to the second layer, W indicates the ginseng between multilayer perceptron adjacent two layers
Number, W and b it is practical be exactly parameter that whole network is trained:
δl=((Wl)Tδl+1)·f'(zl)
4. calculating gradient, Δ W and Δ b indicate the gradient of network training parameter, in back-propagation process each time all with
This parameter updates network:
5. updating weight in multilayer perceptron, η is the learning rate in network, and numberical range general control is in 0 η≤1 <, master
It is used to control the step-length that network parameter training updates:
Wl=Wl-ηΔWl
bl=bl-ηΔbl
Summary of the invention
Part connects the Graphic Design and implementation method of multilayer perceptron, is to construct one in the way of graphic hotsopt
The feedforward neural network that a user needs, the requirement form of network is the graph structure of multilayer perceptron, and one kind is connected based on part
The calculating mode identification method of multilayer perceptron, it is characterised in that following steps:
Step 1: selection training set selects a kind of data set of data set as neural network in board interface, optionally
It is all data set disclosed in UCI that data set is most of, and the present invention has carried the data such as Iris, Sonar, Diabetes, Blood
Collection uses for partially connecting multilayer perceptron model training;By taking Iris data set as an example, data set includes 150 data samples
This, is divided into 3 classes, 50 data of every class, and each data include 4 attributes and 1 label value, can pass through calyx length, calyx
Which kind of in three types width, petal length and 4, petal broadband attribute forecast iris flowers belong to;
Step 1.1: in order to accelerate trained convergence rate, each feature being normalized;It, can according to the sample of selection
With the decline of selection batch gradient, stochastic gradient descent and the decline of mini-batch gradient, because not being optionally large-scale data,
Therefore default uses stochastic gradient descent, only handles a sample every time;Training is had an impact in order to avoid sample order is fixed,
Upset the sequence of sample in data set at random before training sample;
Step 1.2: the characteristic value in sample after pretreatment being put into array X, corresponding label value is successively stored another
In a array Y;It is needed additional for classification task in order to which the error distance calculating allowed between label value and predicted value is more reasonable
Label value is changed into one-hot coding form;If it is the task of recurrence, label value is not processed;
Step 2: tectonic network structure draws the part connection multilayer perceptron for meeting user demand in drawing board interface
Neural network structure;
Step 2.1: in picture interface button regional choice " the multilayer perceptron number of plies " button, then input Multilayer Perception occur
Device number of plies button inputs the number of plies;Then program according to the width at entire picture interface can evenly distribute every layer of spacing, and
The initial layout distribution at picture interface is rationally provided according to the height at picture interface;Part connects each layer of use of multilayer perceptron
One capsule indicates, calculates and saves some vector informations (such as output valve, error term etc.) in capsule structure, and and this
The relevant link information of a capsule;
Step 2.2: neuron number in each layer capsule is set, and clicking capsule then will pop up the corresponding capsule of interface requirements input
Internal neuron number;
Step 2.3: the activation primitive of each layer capsule is set, and capsule is defaulted without activation primitive;Capsule as input is not
Activation processing is done, the activation primitives such as relu, sigmoid, tanh can be used in the capsule of intermediate hidden layers, and output layer capsule swashs
Function living can generally be set as softmax function if it is classification task, not select activation primitive then if it is recurrence task;
Step 2.4: after neuron number determines in the number of plies of part connection multilayer perceptron and every layer of capsule, clicking
" part connects " button, then form part and connect between adjacent two layers capsule;Connection multilayer perceptron in part is generated in interface
Connection between each layer, can be with the sequence of display data stream calculating and the category by clicking the connection of directed edge figure setting section
Property;The mode of part connection can be to be connected at random between two layers of neuron, also can choose the specific of some systems offers
Rule connection;Random partial connection is corresponding to generate a mask matrix, the mask according to the connection Probability p set on directed edge
Matrix is obeyed the Bernoulli that probability is p and is distributed, and only 0 and 1 two element in the matrix, 0 representative disconnects between two nodes
Connection, 1 indicates to keep original connection;
Step 2.5: determining the computation sequence of capsule in the network formed, multilayer sense is mainly connected according to generating portion
Know the dependence list generated in device graphic procedure, determines computation sequence using topological sorting;
Step 3: execution part connects the calculating of multilayer perceptron, the training and test of complete paired data collection;
Step 3.1: executing forward calculation;Unlike the multilayer perceptron calculation of full type of attachment in operation
The mask matrix with weight same shape is used, the connection shape of the intermediate node of two-tier network is controlled by mask matrix
State;W and b indicates the weight and biasing in part connection multilayer perceptron between adjacent two layers, both part connection multilayer
Perceptron needs the parameter of training;Z indicates that one layer of output valve is weighted the knot of summation as the input data of current layer
Fruit, a indicate that activation value of the z after activation primitive calculating, subscript indicate in part connection multilayer perceptron where element
Layer, the forward calculation of the multilayer perceptron connected entirely is as follows:
zl+1=Wlal+bl, al+1=f (zl+1)
And weight matrix is first done corresponding element multiplication with corresponding mask matrix by the multilayer perceptron needs that part connects
(Hadamard product);Mask matrix used in the mask matrix and dropconnect method of part connection multilayer perceptron is not
Same place is that the mask matrix generated in part connection multilayer perceptron is all identical in training and test, and
Mask matrix data of every one batch of training in the training stage will generate a mask at random again in dropconnect
Matrix, and do not use mask matrix in test phase, will also do scaling processing to the parameter in network accordingly, and part
Connection multilayer perceptron does not need to do the scaling processing of parameter in test phase;Here mask matrix is indicated with M, Hadamard product
With indicating, the forward calculation that part connects multilayer perceptron is as follows, and subscript indicates that these parameters connect multilayer in part
Which layer of perceptron:
zl+1=(Ml·Wl)al+bl,al+1=f (zl+1)
Step 3.2: executing retrospectively calculate;The output capsule in network is found, i.e. the last one capsule in calculating figure, utilized
The predicted value of output and the true value of corresponding label first find out the error amount of the last layer, the last layer nlIt indicates, y is indicated
The true value of data sample,What is actually indicated is predicted value of the network to sample, and the process for calculating δ in fact exactly will be last
Overall error is shared in network on each node, and the part using square error as loss function connects multilayer perceptron
The retrospectively calculate formula of the last layer is as follows:
Then it is calculated always in reverse order from layer second from the bottom to the second layer, W and b indicate part connection perceptron adjacent two layers
Between parameter, W and b it is practical be exactly required for whole network training parameter, calculation formula are as follows:
δl=((Wl)Tδl+1)·f'(zl)
Step 3.3: calculating gradient, Δ W and Δ b indicate the gradient of network training parameter, in back-propagation process each time
In be all according to updating network with this parameter, the calculating of gradient is as multilayer perceptron:
Step 3.4: undated parameter value needs exist for again using the upper mask matrix generated before, and the calculating of update is public
Formula is as follows, learning rate of the η for network, η≤1 learning rate general 0 <, for controlling the step-length of network training parameter update:
Wl=Wl-η(Ml·ΔWl)
bl=bl-ηΔbl
Step 3.5: after the training for completing part connection multilayer perceptron, the generalization ability of network is measured on test set,
Calculating process on test set is consistent with the forward calculation in training process, finally exports trained network on test set
Accuracy.
Compared with existing deep learning frame, the present invention is had the effect that
1) it builds neural network model no longer to need since original programming, but passes through shape patterned on interface
The neural network that formula construction needs;
2) part that realizes between layers of multilayer perceptron connects, and can effectively reject full connection multilayer perceptron
Some parameters of middle redundancy.
Detailed description of the invention
Fig. 1 is that the part that present invention capsule constructs connects multilayer perceptron in the effect picture at picture interface.
Fig. 2 is the flow chart that the present invention is implemented.
Specific embodiment
The present invention is explained and illustrated and the multilayer perceptron of more various connection ratios and full type of attachment below
Classifying quality on different data sets.
The multilayer perceptron of graph-based part connection can be used in the present invention, forms an artificial neuron in a short time
Network can be used for intuitively showing the classifying quality of neural network, specific embodiment party in the teaching and experiment of pattern-recognition
Case are as follows:
Step 1: a training set of use is somebody's turn to do from the diabetes data set downloaded in UCI data set on internet
A total of 768 data of data set, feature have pregnancy number, blood glucose, blood pressure, stratum corneum lipids, insulin, BMI body quality
Index, diabetes genetic function, the age, as a result.Wherein the result is that the feature to be predicted, the sample comprising two categories, 0 represents
Diabetes are not suffered from, 1 represents with diabetes.In 768 data points, 500 are marked as 0,268 and are marked as 1.
When selecting the data set and testing, this graphical system will indicate that the 0 of label value, 1 changes into when selecting the data set automatically
One-hot form.The ratio of 7:3 is used for the division of training set and test set.
Step 2: the graphic button provided using front-end interface is formed comprising an input layer, a hidden layer and one
The artificial neural network structure of the multilayer perceptron of output layer.Firstly the need of the capsule figure for clicking different piece in front-end interface
Shape, be arranged in pop-up the capsule expression this layer network in neuron number and this layer in each neuron adopted
Activation primitive.
Step 2.1: in this implementation, since training dataset has eight training characteristics, therefore the mind of input capsule being set
Should be consistent with feature quantity through first number, 8 are set as, activation primitive is not used to input capsule.
Step 2.2: the neuron number that the capsule of intermediate hidden layers is arranged is 100, and activation primitive is selected by combobox
For relu function.
Step 2.3: since label is converted to the one-hot form of expression and is two classification problems, therefore last
The number for exporting setting neuron in capsule is 2, and activation primitive is set as softmax function.
Step 2.4: clicking the ratio partially connected between directed edge figure setting adjacent two layers.
Step 3: in some hyper parameters of the sidebar setting network training of drawing board.It is 0.001 that learning rate, which is arranged, task setting
For classification, training method is stochastic gradient descent, clicks the training that training button starts network.Show that training is every at the end of training
The corresponding accuracy of one wheel and loss curve graph, print display for the accuracy of test set.
The present invention compares 6 kinds of random connection ratios in embodiment, and is same by the connection ratio setting between each layer
Kind, be respectively 0.5,0.6,0.7,0.8,0.9 and for 1 full connection.In order to be comparable experimental result, to training set
Identical cutting is used with the data of test set.It is as shown in table 1 using the test accuracy of same division to the data set:
1 part of table connects multilayer perceptron to the classification performance of UCI Diabetes
Connection ratio | 0.5 | 0.6 | 0.7 | 0.8 | 0.9 | 1.0 |
Test accuracy | 74.63% | 74.85% | 74.98% | 74.89% | 75.06% | 74.68% |
And it also can exceed that in such a way that part connects on the small-sized multilayer perceptron for handling small data set and connect entirely
The effect connect.Classified using the implementation steps of above embodiment to the Iris data set in UCI, since the difference of data set is right
Network structure is adjusted, and the number of nodes of input layer is 4, and the number of nodes of intermediate hidden layers is 50, and the number of nodes of output layer is 3,
Keeping in other parameters and embodiment, the classifying quality on test set is as shown in table 2:
2 part of table connects multilayer perceptron to the classification performance of UCIIris
Connection ratio | 0.5 | 0.6 | 0.7 | 0.8 | 0.9 | 1.0 |
Test accuracy | 94.67% | 95.78% | 94.67% | 94.44% | 93.78% | 92.89% |
Claims (1)
1. a kind of Graphic Design and implementation method of part connection multilayer perceptron, which is characterized in that this method includes as follows
Step:
Step 1: selection training set selects a kind of data set of data set as neural network, optional data in board interface
It is all data set disclosed in UCI that collection is most of, the present invention carried the data sets such as Iris, Sonar, Diabetes, Blood for
Part connection multilayer perceptron model training uses;By taking Iris data set as an example, data set includes 150 data samples, is divided into 3
Class, 50 data of every class, each data include 4 attributes and 1 label value, can pass through calyx length, calyx width, petal
Which kind of in three types length and 4, petal broadband attribute forecast iris flowers belong to;
Step 1.1: in order to accelerate trained convergence rate, each feature being normalized;According to the sample of selection, Ke Yixuan
Batch gradient decline, stochastic gradient descent and the decline of mini-batch gradient are selected, because not being optionally large-scale data, thus it is silent
Recognize using stochastic gradient descent, only handles a sample every time;Training is had an impact in order to avoid sample order is fixed, is being instructed
Upset the sequence of sample in data set before practicing sample at random;
Step 1.2: the characteristic value in sample after pretreatment being put into array X, corresponding label value successively stores another number
In group Y;For classification task, in order to which the error distance calculating allowed between label value and predicted value is more reasonable, needing additionally will mark
Label value changes into one-hot coding form;If it is the task of recurrence, label value is not processed;
Step 2: tectonic network structure draws the nerve for meeting the part connection multilayer perceptron of user demand in drawing board interface
Network structure;
Step 2.1: in picture interface button regional choice " the multilayer perceptron number of plies " button, then occurring inputting multilayer perceptron layer
Number button, inputs the number of plies;Then program according to the width at entire picture interface can evenly distribute every layer of spacing, and according to
The height at picture interface rationally provides the initial layout distribution at picture interface;Each layer of part connection multilayer perceptron is with one
Capsule indicate, calculate and save some vector informations (such as output valve, error term etc.) in capsule structure, and with this glue
The relevant link information of capsule;
Step 2.2: neuron number in each layer capsule is set, and clicking capsule then will pop up the corresponding capsule of interface requirements input
Neuron number;
Step 2.3: the activation primitive of each layer capsule is set, and capsule is defaulted without activation primitive;Capsule as input does not do sharp
The activation primitives such as relu, sigmoid, tanh, the activation letter of output layer capsule can be used in processing living, the capsule of intermediate hidden layers
Number can generally be set as softmax function if it is classification task, not select activation primitive then if it is recurrence task;
Step 2.4: after neuron number determines in the number of plies of part connection multilayer perceptron and every layer of capsule, clicking " part
Connection " button then forms part and connects between adjacent two layers capsule;In interface generate part connection each layer of multilayer perceptron it
Between connection, can with display data stream calculate sequence and by click directed edge figure setting section connection attribute;Part
The mode of connection can be to be connected at random between two layers of neuron, also be can choose the specific rule that some systems provide and is connected
It connects;Random partial connection is corresponding to generate a mask matrix, mask matrix clothes according to the connection Probability p set on directed edge
It is distributed from the Bernoulli that probability is p, only 0 and 1 two element in the matrix, 0 represents the company disconnected between two nodes
It connects, 1 indicates to keep original connection;
Step 2.5: determining the computation sequence of capsule in the network formed, multilayer perceptron is mainly connected according to generating portion
The dependence list generated in graphic procedure, determines computation sequence using topological sorting;
Step 3: execution part connects the calculating of multilayer perceptron, the training and test of complete paired data collection;
Step 3.1: executing forward calculation;Benefit is wanted in operation unlike the multilayer perceptron calculation of full type of attachment
The mask matrix with weight same shape is used, the connection status of the intermediate node of two-tier network is controlled by mask matrix;W and
B indicates the weight and biasing in part connection multilayer perceptron between adjacent two layers, both part connection multilayer perceptron
Need trained parameter;Z indicate one layer of output valve as the input data of current layer be weighted summation as a result, a table
Show that the activation value after z is calculated by activation primitive, subscript indicate the layer in part connection multilayer perceptron where element, entirely
The forward calculation of the multilayer perceptron of connection is as follows:
zl+1=Wlal+bl, al+1=f (zl+1)
And weight matrix is first done corresponding element multiplication with corresponding mask matrix by the multilayer perceptron needs that part connects
(Hadamard product);Mask matrix used in the mask matrix and dropconnect method of part connection multilayer perceptron is not
Same place is that the mask matrix generated in part connection multilayer perceptron is all identical in training and test, and
Mask matrix data of every one batch of training in the training stage will generate a mask at random again in dropconnect
Matrix, and do not use mask matrix in test phase, will also do scaling processing to the parameter in network accordingly, and part
Connection multilayer perceptron does not need to do the scaling processing of parameter in test phase;Here mask matrix is indicated with M, Hadamard product
With indicating, the forward calculation that part connects multilayer perceptron is as follows, and subscript indicates that these parameters connect multilayer in part
Which layer of perceptron:
zl+1=(Ml·Wl)al+bl,al+1=f (zl+1)
Step 3.2: executing retrospectively calculate;The output capsule in network is found, i.e. the last one capsule in calculating figure, utilizes output
Predicted value and corresponding label true value, first find out the error amount of the last layer, the last layer nlIt indicates, y indicates data
The true value of sample,What is actually indicated is predicted value of the network to sample, and the process for calculating δ is exactly by last total mistake in fact
Into network on each node, the part using square error as loss function connects the last of multilayer perceptron at difference booth
One layer of retrospectively calculate formula is as follows:
Then it is calculated always in reverse order from layer second from the bottom to the second layer, W and b are indicated between part connection perceptron adjacent two layers
Parameter, W and b it is practical be exactly required for whole network training parameter, calculation formula are as follows:
δl=((Wl)Tδl+1)·f'(zl)
Step 3.3: calculating gradient, Δ W and Δ b indicate the gradient of network training parameter, in back-propagation process each time all
Network is updated so that this parameter is foundation, the calculating of gradient is as multilayer perceptron:
Step 3.4: undated parameter value needs exist for again using the upper mask matrix generated before, and the calculation formula of update is such as
Shown in lower, learning rate of the η for network, η≤1 learning rate general 0 <, for controlling the step-length of network training parameter update:
Wl=Wl-η(Ml·ΔWl)
bl=bl-ηΔbl
Step 3.5: after the training for completing part connection multilayer perceptron, the generalization ability of network is measured on test set, is being surveyed
Calculating process on examination collection is consistent with the forward calculation in training process, finally exports trained network on test set just
True rate.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811538198.3A CN109635932A (en) | 2018-12-16 | 2018-12-16 | A kind of Graphic Design and implementation method of part connection multilayer perceptron |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811538198.3A CN109635932A (en) | 2018-12-16 | 2018-12-16 | A kind of Graphic Design and implementation method of part connection multilayer perceptron |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109635932A true CN109635932A (en) | 2019-04-16 |
Family
ID=66074450
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811538198.3A Pending CN109635932A (en) | 2018-12-16 | 2018-12-16 | A kind of Graphic Design and implementation method of part connection multilayer perceptron |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109635932A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111370080A (en) * | 2020-03-05 | 2020-07-03 | 中国工程物理研究院上海激光等离子体研究所 | Radiation temperature inversion method based on artificial neural network algorithm |
CN113256593A (en) * | 2021-06-07 | 2021-08-13 | 四川国路安数据技术有限公司 | Tumor image detection method based on task self-adaptive neural network architecture search |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106022273A (en) * | 2016-05-24 | 2016-10-12 | 华东理工大学 | Handwritten form identification system of BP neural network based on dynamic sample selection strategy |
CN107678746A (en) * | 2017-09-30 | 2018-02-09 | 江西洪都航空工业集团有限责任公司 | A kind of pattern development platform based on ROS |
CN108319456A (en) * | 2018-01-29 | 2018-07-24 | 徐磊 | A kind of development approach for exempting to program deep learning application |
CN108563977A (en) * | 2017-12-18 | 2018-09-21 | 华南理工大学 | A kind of the pedestrian's method for early warning and system of expressway entrance and exit |
CN109002879A (en) * | 2018-07-23 | 2018-12-14 | 济南浪潮高新科技投资发展有限公司 | The visual modeling method and device of neural network model |
-
2018
- 2018-12-16 CN CN201811538198.3A patent/CN109635932A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106022273A (en) * | 2016-05-24 | 2016-10-12 | 华东理工大学 | Handwritten form identification system of BP neural network based on dynamic sample selection strategy |
CN107678746A (en) * | 2017-09-30 | 2018-02-09 | 江西洪都航空工业集团有限责任公司 | A kind of pattern development platform based on ROS |
CN108563977A (en) * | 2017-12-18 | 2018-09-21 | 华南理工大学 | A kind of the pedestrian's method for early warning and system of expressway entrance and exit |
CN108319456A (en) * | 2018-01-29 | 2018-07-24 | 徐磊 | A kind of development approach for exempting to program deep learning application |
CN109002879A (en) * | 2018-07-23 | 2018-12-14 | 济南浪潮高新科技投资发展有限公司 | The visual modeling method and device of neural network model |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111370080A (en) * | 2020-03-05 | 2020-07-03 | 中国工程物理研究院上海激光等离子体研究所 | Radiation temperature inversion method based on artificial neural network algorithm |
CN111370080B (en) * | 2020-03-05 | 2023-12-08 | 中国工程物理研究院上海激光等离子体研究所 | Radiation temperature inversion method based on artificial neural network algorithm |
CN113256593A (en) * | 2021-06-07 | 2021-08-13 | 四川国路安数据技术有限公司 | Tumor image detection method based on task self-adaptive neural network architecture search |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109960759B (en) | Recommendation system click rate prediction method based on deep neural network | |
Alam et al. | A dynamic ensemble learning algorithm for neural networks | |
Guo et al. | Efficient hyperparameter optimization for convolution neural networks in deep learning: A distributed particle swarm optimization approach | |
Wu et al. | Evolving RBF neural networks for rainfall prediction using hybrid particle swarm optimization and genetic algorithm | |
CN111062775B (en) | Recommendation system recall method based on attention mechanism | |
CN110083700A (en) | A kind of enterprise's public sentiment sensibility classification method and system based on convolutional neural networks | |
CN109740154A (en) | A kind of online comment fine granularity sentiment analysis method based on multi-task learning | |
CN107358169A (en) | A kind of facial expression recognizing method and expression recognition device | |
CN106503654A (en) | A kind of face emotion identification method based on the sparse autoencoder network of depth | |
CN109345302A (en) | Machine learning model training method, device, storage medium and computer equipment | |
Tuba et al. | Hybrid seeker optimization algorithm for global optimization | |
WO2019223302A1 (en) | Dress collocation method and system based on attention knowledge extraction, and storage medium | |
CN107341145A (en) | A kind of user feeling analysis method based on deep learning | |
CN106295507A (en) | A kind of gender identification method based on integrated convolutional neural networks | |
CN109582782A (en) | A kind of Text Clustering Method based on Weakly supervised deep learning | |
CN106897404A (en) | A kind of recommendation method and system based on many GRU layers of neutral nets | |
CN106919710A (en) | A kind of dialect sorting technique based on convolutional neural networks | |
CN109766935A (en) | A kind of semisupervised classification method based on hypergraph p-Laplacian figure convolutional neural networks | |
CN107223260A (en) | Method for dynamicalling update grader complexity | |
CN109635932A (en) | A kind of Graphic Design and implementation method of part connection multilayer perceptron | |
Nayak et al. | A comprehensive review and performance analysis of firefly algorithm for artificial neural networks | |
CN108230121A (en) | A kind of product design method based on Recognition with Recurrent Neural Network | |
Lagani et al. | Training convolutional neural networks with competitive hebbian learning approaches | |
Zhang et al. | Deep compression of probabilistic graphical networks | |
CN106127208A (en) | Method and system that multiple objects in image are classified, computer system |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190416 |
|
WD01 | Invention patent application deemed withdrawn after publication |